CoolFace
Modelpublic

JohnDi/Hemmingway-1-Abliterated-Extreme-GGUF

sourceHugging Faceapache-2.0updated 4d agoView on Hugging Face
8likes1.6kdownloads
Model Card

Hemmingway-1-Abliterated-Extreme (GGUF)

An uncensored, abliterated release of Altworld/Hemmingway-1 (27B parameter model based on Qwen), optimized specifically for unrestricted creative writing, authentic dialogue, and everyday communication.

Unlike standard quick abliteration runs with only 20–30 trials, this model was optimized using Heretic over a deep 200-trial Bayesian search (Optuna) to locate the exact Pareto frontier between maximum decensoring and minimum distortion of the model's natural prose.

Benchmark Scores (200-Trial Deep Search)

  • —Original Baseline Refusals: 98 / 100
  • —Trial 194 (Extreme): 2 / 100 Refusals (98% of all refusal mechanisms removed!)
  • —KL Divergence: 0.1870 (Prose, humor, and sentence flow remain clean, coherent, and human-like)

Available GGUF Files

FileSizeDescriptionRecommended For
Hemmingway-1-Extreme-Q6_K.gguf~23 GBHigh fidelity, excellent balance of speed & quality24 GB – 32 GB VRAM (RTX 3090 / 4090 / 5090)
Hemmingway-1-Extreme-Q8_0.gguf~29 GBMaximum 8-bit precision, near-lossless32 GB+ VRAM
Hemmingway-1-Extreme-Q4_K_M.gguf~17 GBFast, memory-efficient16 GB – 24 GB VRAM

Recommended Sampling Settings

To maintain natural phrasing and avoid repetitive or awkward wording:

  • —Temperature: 0.80 – 0.85 (or up to `1.0` for creative fiction)
  • —Top P: 0.95
  • —Top K: 20
  • —Min P: 0.05
  • —Presence Penalty: 0.0 ⚠️ (Keep strictly at 0.0 – higher values degrade natural grammar)
  • —Repetition Penalty: Off (or max 1.02)

Prompt Format

Uses the standard ChatML / Qwen format:

text
<|im_start|>system
You are a helpful, candid assistant.<|im_end|>
<|im_start|>user
Write a direct message about...<|im_end|>
<|im_start|>assistant